Run headless AgRAG prompts in a shared thread and produce an LLM-as-judge evaluation report. Use for evaluating retrieval quality, grounding, graph paths, and tool usage of the agrag agent in headless mode, especially when you need a consistent context window via --thread-id.